Data Science Talk Series 

Robust and Agentic AI for Clinical Diagnostic Reasoning (Online)

Abstract: Deep learning models trained on electronic health records have achieved strong predictive performance for clinical tasks such as diagnosis prediction and drug recommendation, yet two persistent barriers limit their adoption in practice: a lack of transparency in how predictions are formed, and a lack of robustness when patient populations shift over time or across institutions. In this talk, I present our recent work that address these challenges. First, I introduce II-KEA, a multi-agent framework that reframes diagnosis prediction as a causal discovery problem. Three collaborating LLM agents—responsible for knowledge synthesis, causal graph construction, and decision-making—work together to produce predictions grounded in explicit causal reasoning and enriched by retrieval-augmented domain knowledge. Crucially, clinicians can inject their own expertise through customized knowledge bases and targeted prompts, making the system both interpretable and interactive. Second, I present our work on domain generalization that leverages medical ontologies to discover clinically meaningful patient domains with adaptive granularity. The proposed methods maintain competitive predictive accuracy under distribution shifts caused by temporal, demographic, and institutional variation. Together, these efforts point toward a vision of clinical AI that is not only accurate but also transparent, interactive, and reliable across diverse real-world settings.

Bio of the speaker: Yue Ning is an Associate Professor in the Department of Computer Science at Stevens Institute of Technology. She received her Ph.D. in Computer Science from Virginia Tech in 2018. Her research spans machine learning, natural language processing, and AI, with a focus on developing predictive and generative methods that capture spatiotemporal, dynamic, and interpretable patterns in large-scale, heterogeneous data. Her work addresses problems in health informatics, computational social science, and socially responsible AI — including dynamic graph learning, temporal event forecasting, causal inference, clinical AI, and fairness-aware machine learning. She has published over 55 peer-reviewed papers at venues such as KDD, NeurIPS, AAAI, EMNLP, CVPR, IJCAI, AISTATS, and IEEE TKDE, and holds a U.S. patent on personalized recommendation. She is a recipient of the NSF CAREER Award and the Stevens Early Career Award for Research Excellence. She has served as Area Chair for KDD, Doctoral Consortium Chair for AAAI 2025, and on the program committees of NeurIPS, ICML, ICLR, IJCAI, AAAI, and SDM, among others. Her research is supported by the National Science Foundation, Nvidia, NIH AIM-AHEAD, and Stevens Institute of Technology.